Drafts Sometimes, Review Never
Where generated text belongs in a documentation workflow, and where it does not.
We use generated drafts and we publish nothing without a named human maintainer signing it off. Both halves of that matter.
What It Is Genuinely Good At
First drafts of repetitive reference pages where the schema already exists. Rewriting a support reply into an article shape. Generating the boring half of an example in six languages. All of these are real time savings on work that is structurally tedious.
Where It Fails Quietly
Plausible-sounding parameters that do not exist. Confidently wrong default values. Code examples that would compile but do the wrong thing. The failure mode is not obvious nonsense — it is text that reads exactly like documentation and is subtly incorrect, which is the worst possible failure for a reference.
Which Is Why Review Is Not Optional
Every published page has a named maintainer accountable for its accuracy, and that name is on the page. Nobody signs off a reference page without checking each parameter against the actual schema, whether a person or a model wrote the first draft.
What We Tell Users
The docs FAQ answers this directly, because people increasingly and reasonably want to know. Being straightforward about it is cheaper than being caught being vague.
